Some remarks on two interval-arithmetic modifications of the Newton method (Q1192015)

The paper is concerned with the interval Newton algorithm and that modified version of it where the Jacobian is kept constant. Let \(I\) be the set of real compact intervals, \(f:\mathbb{R}^ n\to\mathbb{R}^ n\) be the function to be solved, and \(X_ 0\in I^ n\) the starting vector. For \(C\in I^{n\times n}\), \(b\in I^ n\) let \(\text{IGA}(C,b)\) be the result of applying the interval Gaussian elimination to the equation \(Cx=b\), and \(\text{IGA}(C)\) be the matrix product connected with the product representation of \(\text{IGA}(C,b)\). Let finally \(A=| Id- (\text{IGA}(f'(X_ 0)))f'(X_ 0)|\) and \(B=[d(\text{IGA}(f'(X_ 0)))]| f'(X_ 0)|\). Firstly, a question raised by \textit{G. Alefeld} [SIAM J. Numer. Anal. 21, 363-372 (1984; Zbl 0536.65026)] is answered confirming that the 2 algorithms will deliver empty intervals after a finite number of iterations, if \(\text{IGA}(f'(X_ 0))\) exists, \(X_ 0\) contains no zero of \(f\), and the spectral radius of \(B\) is smaller than 1. Secondly, a few relations between \(A\) and \(B\) are derived.